This work introduces a novel application of chemometric-based preprocessing and machine learning for fire debris analysis, which is crucial in origin and cause investigations. The most technically demanding and interpretative aspect of fire debris analysis is the qualitative classification of GC-MS data using pattern matching, which can be performed by artificial intelligence to support human analysts. Three different methods for preprocessing GC-MS data for machine learning, each requiring varying levels of analyst input, were developed using chromatographic software for feature extraction and data export. These methods were evaluated alongside several machine and deep learning models to classify fire debris and liquid samples containing self-heating fatty acids. The dataset comprised 310 samples (153 positive and 157 negative) generated from spontaneous heating experiments, neat exemplars, forensic casework, and pyrolyzed substrates. Models trained on each preprocessing method were evaluated using repeated 2-fold and Monte Carlo cross-validation across multiple training/testing splits. Within the scope of this data set and preprocessing methods, naive bayes, random forest, and gradient boosting performed best across 2-fold evaluations, with mean accuracies of 100 %, 99.90 %, and 99.65 % for the three preprocessing methods. This pilot study demonstrates a novel, chemometric workflow for fatty acid classification and establishes options for extending machine learning to more complicated fire debris tasks such as ignitable liquid residue analysis. The results imply that machine learning has the potential to enhance fire debris analysis by improving accuracy and analytical efficiency by streamlining routine classification tasks, allowing laboratories to allocate expert effort more effectively and reduce turnaround time.
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号